Uncertainty estimation for deep learning-based pectoral muscle segmentation via Monte Carlo dropout

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Deep Learning models can fail post-deployment, necessitating methods to detect inadequate predictions.
• Accurate pectoral muscle segmentation in mammograms is vital for breast cancer screening.

Purpose of the Study:

• To evaluate Monte Carlo (MC) dropout and a novel uncertainty metric (UM) for identifying unacceptable pectoral muscle segmentations.
• To assess the reliability of the UM in flagging segmentation errors in mammograms.

Main Methods:

• Pectoral muscle segmentation using a modified ResNet18 convolutional neural network with MC dropout enabled at inference.
• Generating 50 segmentations per mammogram to calculate mean segmentation and standard deviation for uncertainty estimation.
• Validating the UM against the Dice Similarity Coefficient (DSC) and using ROC-AUC analysis.

Main Results:

• MC dropout improved segmentation performance (DSC = 0.95 ± 0.07).
• A strong negative correlation (r = -0.76) was found between UM and DSC.
• The UM achieved high discriminatory power for unacceptable segmentations (AUC = 0.98).

Conclusions:

• MC dropout and the proposed UM effectively flag unacceptable pectoral muscle segmentations in mammograms.
• The UM demonstrates excellent discriminatory power, enhancing the reliability of AI in medical imaging.